Research And Implementation Of Feature Extraction Algorithm For Ultrasonic Echo Signal In Imaging Logging

Ultrasonic imaging logging is a promising technology which is widely used in oil and gas exploration since it can accomplish full borehole detection,provide intuitive logging results and assist in qualitatively analyzing formation structure.This technology's principle is that the instrument sends pulsed ultrasonic signals to borehole walls,and detect the arrival time and the amplitude of the corresponding reflected echo signals.According to the arrival time and the amplitude at different positions,the corresponding borehole wall image can be drawn.Therefore,it is considerably essential to accurately determine the arrival time of the ultrasonic echo signal for drawing accurate and clear logging image.This thesis aims at the accurately determining arrival time of ultrasonic echo signal,and designs corresponding detection algorithms.Meanwhile,experiments have been conducted to analyze proposed methods' performances.The research of the thesis mainly consists of the following aspects:1.Features of ultrasonic echo signals have been analyzed.Considering the timefrequency,statistical characteristics and stationarity analysis of echo signals,a new arrival time determination approach is proposed based on wavelet transform and Bayesian information criterion.By experiments based on echo signals collected in a testing well,the method is demonstrated to have satisfactory anti-noise robustness and little detection deviation.2.Aiming at overlapping of ultrasonic echo signal in cased hole detection,based on deep learning,a method is proposed termed Wave Transform Network.The network consists of two parts: separation network(S-net)and detection network(D-net),which are used for separation of multiple signals and arrival time detection of separated echo signals so as to achieve end-to-end multiple echo signal time detection.The ultrasonic signals collected from experimental wells are tested by this method,and the results show that the method has good separation and detection performance.3.Due to the requirements of low power consumption and high computing power for logging tools,more stringent restrictions are put forward for the deployment and implementation of the algorithm.Therefore,the thesis preliminarily explores the deployment method of neural network in embedded devices,and proposes a method for ultrasonic echo arrival time detection based on quantized neural network.Through the FINN framework,the network is deployed in FPGA.The test results show that under the premise of ensuring the inspection accuracy,the scale of the network is greatly compressed compared with the conventional structure of the neural network,and it shows a good application prospect.